[Request]: mtp layers weights #47
Description
Proposal
Introduction
Hi, enjoying the model and wanted to ask a question - there are some discussions in the community section on the huggingface model page, but i found no answer there.
Model has mtp layers inside of the configuration, but weights of these mtp modules are not published(they have garbage data inside, intentionally, i suppose). Are you planning on publishing mtp layers? If no, maybe you can propose the model, that can be used for speculative decoding using eagle?
Rational
Implementation of this feature will help the following usecase:
- GLM recently fixed mtp layers in their existing models, so vllm and sglang will allow users to launch speculative decoding, increasing speed. Speculative decoding for minimax-m2 will help model to be more competitive in performance to upcomming glm-4.6-air